---
license: cc-by-nc-4.0
task_categories:
- robotics
- imitation-learning
- image-to-image
tags:
- vitacdreamer
- univtac
- tactile
- visual-tactile
- act
- robotics
---

# ViTacDreamer Lift Bottle Assets

This dataset repository contains the processed assets used for the `lift_bottle/default-50` ViTacDreamer, ACT baseline, and ViTacACT experiments.

## Contents

- `data/act_processed/default-50/`: processed ACT training episodes generated from UniVTAC `lift_bottle/clean`.
- `data/act_processed/default-50/vitacdreamer_features_stage2/`: cached frozen ViTacDreamer features for accelerated ViTacACT training.
- `checkpoints/vitacdreamer/lift_bottle_stage1/`: ViTacDreamer Stage 1 visual-tactile contrastive checkpoint.
- `checkpoints/vitacdreamer/lift_bottle_stage2/`: ViTacDreamer Stage 2 cVAE reconstruction checkpoint.
- `checkpoints/act_lift_bottle/default-50/train_config/`: ACT baseline checkpoints.
- `checkpoints/act_lift_bottle/default-50/train_config_vitacdreamer/`: earlier ViTacACT intermediate checkpoints.
- `checkpoints/act_lift_bottle/default-50/train_config_vitacdreamer_cached/`: final cached ViTacACT checkpoints.
- `configs/`: release-ready ACT, ViTacACT, and cached ViTacACT training configs with dataset-local paths.
- `scripts/precompute_vitacdreamer_features.py`: script used to precompute ViTacDreamer features for cached ViTacACT training.
- `scripts/upload_to_hf.py`: resumable uploader for pushing this folder to a Hugging Face dataset repository.

## Key Results

- ACT baseline best validation loss: `0.107912`.
- Cached ViTacACT best validation loss: `0.093634` at epoch 78.

## Release Scope

- Processed ACT dataset episodes: `50`
- Cached ViTacDreamer feature files: `50`
- ViTacDreamer checkpoints: Stage 1 and Stage 2 best checkpoints
- Policy checkpoints: ACT baseline, online ViTacACT intermediate checkpoints, cached ViTacACT best/last checkpoints

## Usage Notes

- The configs in `configs/` assume you run training or evaluation commands from the dataset root.
- `tactile_ckpt` is set to `null` in the release configs because the original private tactile encoder checkpoint is not included in this dataset release.
- Cached ViTacACT training uses `data/act_processed/default-50/vitacdreamer_features_stage2/` and does not need to recompute frozen ViTacDreamer features.
- If you want to regenerate cached features, use `scripts/precompute_vitacdreamer_features.py` together with the released Stage 2 checkpoint.

## Notes

The cached ViTacACT setup precomputes ViTacDreamer features first, then trains ACT using cached feature tokens to avoid recomputing the frozen ViTacDreamer encoder at every policy-training step.

The assets are intended for research use and reproducibility of the `lift_bottle/default-50` experiments.